Connect your own agent
Connect any MCP server on your machine to Adunem with a manifest, then register it and keep it connected.
If you have your own agent, a script, a model setup or a tool you built, you can connect it as long as it speaks MCP (the Model Context Protocol). Adunem reads the tools your MCP server offers, shows them on the agent's page, and carries calls to it. Your code, your model and your API keys stay on your machine.
You need the adunem command line tool (see Install the CLI) and an MCP server that runs from a folder on your computer.
Connect it
Sign in
Approve this computer in your browser. You do this once.
adunem loginCreate the manifest
In the agent's folder, run init. It asks for a name, a description, how the server starts, and writes agentmesh.yaml. It never overwrites an existing one.
cd my-agent
adunem initAdd --yes to accept the defaults, or pass answers as options such as --name, --transport and --command.
Register it
This creates the agent in your account and saves its identity in .agentmesh/identity.json in the folder. The file holds no secrets. It is safe to run again.
adunem registerConnect it
This starts your server, connects it to Adunem and keeps it connected until you press Ctrl-C. It reconnects by itself if the network drops.
adunem connectThe agent now appears under Agents as Online, with the tools your server reported. Open it and use Try it or the Playground: see Try an agent.
The Connect page does all of this in one go. Choose My own agent, copy the command it gives you and run it in the agent's folder. It finds the start command (or uses your agentmesh.yaml), registers the agent, keeps it online from login on, and can add it to a space.
To keep the agent connected after you close the terminal, see Stay online.
The manifest
agentmesh.yaml tells Adunem how to run your server. This is a typical one:
schema_version: 1
name: adapter-agent
description: Plain Python functions exposed as MCP tools
mcp:
transport: stdio
command: python
args: [server.py]
cwd: .
environment:
pass_through: []
execution:
default_timeout_seconds: 60
tool_timeouts: {}| Key | What it does |
|---|---|
name | Lowercase letters, digits, - and _. This is the agent's name in the dashboard. |
description | One line shown on the agent's page. |
mcp | How to reach your MCP server: stdio or streamable_http (see below). |
environment.pass_through | The names of environment variables your server may see. Only these are forwarded; their values are read on your machine and never uploaded. |
execution | How long a call may run: default_timeout_seconds for all tools, tool_timeouts for particular ones. Timeouts are 1 to 1800 seconds. |
Unknown keys are rejected, so a typo shows up straight away. See The manifest for every key.
Stdio or Streamable HTTP
Adunem starts your server as a child process, talks to it over its standard input and output, and stops it when you disconnect. This is the simplest option and what adunem init offers by default.
mcp:
transport: stdio
command: python
args: [server.py]
cwd: .The process runs with your user's permissions. The working directory is not a sandbox, so only connect agents you would run yourself.
Tool names that do something special
chat: if your server has a tool calledchat, the agent gets a Chat section in the dashboard,adunem chatcan talk to it, and people you share it with can chat too. A chat tool takes amessageand optionally recent turns ashistory.browse: an agent with abrowsetool can work in web pages through the browser extension.
Tool inputs come from your server's own schema. In the Playground they become a form, and a fixed choice (for example a Literal type in Python) becomes a dropdown. Raise a normal error for bad input so the caller sees your message: an unexpected crash hides its message from the caller.
A complete example
A small agent with one tool, in two files. It needs only uv, which
the Adunem installer already set up; uv fetches the MCP SDK when the agent starts.
schema_version: 1
name: my-tools
description: Adds up, averages and compares numbers
mcp:
transport: stdio
command: uv
args: [run, --no-project, --with, "mcp>=2", server.py]Put both files in a folder, then run adunem register and adunem connect there. In the
Playground, calculate shows a dropdown for operation and a list field for numbers.
To add a tool, add another function with @mcp.tool(): its name, inputs and description come
from the function's signature and docstring.